specification tests for the multinomial logit model · october1981 commentswelcome...
TRANSCRIPT
department
of economics
SPECIFICATION TESTS FOR THE MULTINOMIAL LOGIT MODEL
Jerry A. HausmanDaniel McFadden
Number 292 October 1981
massachusetts
institute of
technology
50 memorial drive
Cambridge, mass. 02139
Comments welcome
SPECIFICATION TESTS FOR THE MULTINOMIAL LOGIT MODEL
Jerry A. HausmanDaniel McFadden
Number 292 October 1981
Jeff Dubin, Whitney Newey, and John Rust providedresearch assistance. The NSF and DOE provided researchsupport
.
October 1981
Comments Welcome
SPECIFICATION TESTS FOR THE MULTINOMIAL LOGIT MODEL
by
JERRY HAUSMAN
AND
DANIEL MCFADDEN
Jeff Dubin, Whitney Newey, and John Rust provided research
assistance. The NSF and DOE provided research support.
Digitized by the Internet Archive
in 2011 with funding from
Boston Library Consortium IVIember Libraries
http://www.archive.org/details/specificationtes00haus2
SPECIFICATION TESTS FOR THE MULTINOMIAL LOGIT MODEL
by
JERRY HAUSMAN AND DANIEL MCFADDEN
Discrete choice models are now used in a wide variety of situations in
applied econometrics.^ By far the model specification which is used most
often is the multinomial logit model, McFadden (1974). The multinomial logit
model provides a convenient closed form for the underlying choice
probabilities without any requirement of multivariate integration.
Therefore, choice situations characterized by many alternatives can be
treated in a computationally convenient manner. Furthermore, the likelihood
function for the multinomial logit specification is globally concave which
also eases the computational burden. The ease of computation and the
existence of a number of computer programs has led to the many applications
of the logit model. Yet it is widely known that a potentially important
drawback of the multinomial logit model is the independence from irrelevant
alternatives property. This property states that the ratio of the
probabilities of choosing any two alternatives is independent of the
attributes of any other alternative in the choice set .^ Debreu (1960) was
among the first economists to discuss the implausibility of the independence
from irrelevant alternatives assumption. Basically, no provision is made for
^McFadden (1981) provides references to many of their uses.
^A "universal" logit model avoids the independence from irrelevantalternatives property while maintaining the multinomial logit form by making
each ratio of probabiities a function of attributes of all alternatives,
McFadden (1981). It is difficult however to give an economic interpretation
of this model other than as a flexible approximation to a general functional
form.
different degrees of substitutability or corapLimentarity among the choices.
While most analysts recognize the implications of the independence of
irrelevant alternatives property, it has remained basically a maintained
assumption in applications.
The multinomial probit model does provide an alternative specification
for discrete choice models without any need for the independence of
irrelevant alternatives assumption, Hausnian and Wise (1978). Furthermore, a
test of the 'covariance' probit specification versus the 'independent' probit
specification which is very similar to the logit specification does provide a
test for the independence from irrelevant alternatives assumption. But use
of the multinomial probit model has been limited due to the requirement that
multivariate normal integrals must be evaluated to estimate the unknown
parameters. Thus, the multinomial probit model does not provide a convenient
specification test for the multinomial logit model because of its
complexity.
In this paper we provide two sets of computationally convenient
specification tests for the multinomial logit model. The first test is an
application of the Hausman (1978) specification test procedure. The basic
idea for the test here is to test the reverse implication of the independence
from irrelevant alternatives property. The usual implication is to note that
3
if two choices exist, say car and bus in a transportation choice application,
that addition of a third choice, subway, will not change the ratio of
probabilities of the initial two choices. Our test here is based on
eliminating one or more alternatives from the choice set to see if underlying
choice behavior from the restricted choice set obeys the independence from
irrelevant alternatives property. We estimate the unknown parameters from
both the unrestricted and restricted choice sets. If the parameter estimates
are here approximately the same, then we do not reject the multinomial logit
specification. The test statistic is easy to compute since it only requires
computation of a quadratic form which involves the difference of the
parameter estimates and the differences of the estimated covariance matrices.
Thus, existing logit computer programs provide all the necessary input to the
test.
The second set of specification tests that we propose are based on more
classical test procedures. We consider a generalization of the multinomial
logit model which is called the nested logit model, McFadden (1981). Since
the multinomial logit model is a special case of the more general model when
a given parameter equals one, classical test procedures such as the Wald,
likelihood ratio, and Lagrange multiplier tests can be used. Of course, we
have added the requirement of the specification of an alternative model to
test the original model specification. Maximum likelihood estimation of the
nested logit model is considerably more difficult than for the multinomial
logit model. Thus we base our test procedures on one-step asymptotically
equivalent estimators to maximum likelihood. Still, new programming is
4
required since existing multinomial logit programs do not provide asymptotic
covariance matrix estimates for the one-step estimators.
We then proceed to compare the two sets of specification test procedures
for an example. We find rather unexpected results. First despite a sample
size of 1000, the asymptotically equivalent classical tests differ markedly
in their operating characteristics. The exact size of the Wald test is 15-
25% larger than the nominal size while the exact size of the Lagrange
multiplier test is approximately 15-30% smaller than the nominal size. The
exact size of the likelihood ratio test is quite close to its nominal size.
Furthermore, the power of the Wald test is significantly greater than the
other two classical tests. This result holds also when the tests are
corrected for size. Thus, the Lagrange multiplier test which is based in
this application on inconsistent estimates is distinctly inferior to the Wald
test which is based on asymptotically efficient estimates. Perhaps, more
surprising, we find the power of one Hausman test to be comparable to that of
the Wald test, even though in our example this Wald test is based on the
correct alternative model. The exact size of this Hausman test is within .1%
of the nominal size. When it is compared to the size corrected Wald test, it
proves to be superior. Thus the often quoted asymptotic power results for
local departures from the null hypothesis do not provide a reliable guide to
the exact performance of our specification tests in the example despite the
relatively large sample size and small departures from the multinomial logit
model
.
The plan of the paper is as follows. In the next section we derive the
Hausman-type specification test for the multinomial logit model. The
5
distribution theory as well as computational considerations are discussed.
In Section 3 we apply the specification test to an actual choice situation.
In the next section we derive the classical tests from the nested logit
model. We apply the tests to the same data as we did for the Hausraan test.
All four tests lead to a decisive rejection of the multinomial logit model in
this application. In Section 4 we calculate the exact size and power for
the two sets of specification tests for an example. Lastly, in the
conclusion we report some additional empirical results and discuss some
further considerations for the test procedures.
T . A Hausman-Type Test of the IIA Property
A widely used functional form for discrete choice probabilities is the
multinomial logit (MNL) model
z.g z.g
P(i|z, C, e) = e ^ / .Z e J(1.1)
where
C = {l,...,j} is a finite choice set
i,j = alternatives in C
z. = a K-vector of explanatory variables describing the attributes
of alternatives j and/or the characteristics of the decision-
maker which affect the desirability of alternative j.
z = (zjL,...,z ) the attributes of C
B = a K-vector of taste parameters
P(i|z, C, 3) = the probability that a randomly selected decision-
maker when faced with choice set C with attributes z
will choose i.
The MNL model has a necessary and sufficient characterization, termed
independence from irrelevant alternatives (IIA), that the ratio of the
probabilities of choosing any two alternatives is independent of the
attributes or the availability of a third alternative, or
P(i|z, C, 3) = P(i|z, A, 3) P(A|z, C, 3) (1.2)
where i e A c_ C and
P(A|z, C, 3) = .2 P(j Iz, C, 3) (1.3)
7
This property greatly facilitates estimation and forecasting because it
implies the model can be estimated from data on binomial choices, or by
restricting attention to choice within a limited subset of the full choice
set. On the other hand, this property severely restricts the flexibility of
the functional form, forcing equal cross-elasticities of the probabilities of
choosing various alternatives with respect to an attribute of one
alternative. Further discussion of the IIA property and conditions under
which it is likely to be true or false is given in Doraencich and McFadden
(1975) McFadden, Tye , and Train (1976), and Hausraan-Wise (1978). The
McFadden, Tye and Train paper suggests that the MNL specification be tested
by comparing parameter estimates obtained from choice data from the full
choice set with estimates obtained from conditional choice data from a
restricted choice set. Here we develop an asymptotic test statistic for this
comparison, using the approach to specification tests introduced by Hausman
(1978).
Consider a random sample with observations n=l,...,N. Let z be the
attributes of C for case n, and define S. = 1 if case n chooses i andmS. =0 otherwise. The log Likelihood of the sample ism
N
L^(e) =-J: L, -^r S. In P(i|z", C, 3) (1.4)
C N n=l leC m
We first review the asymptotic properties of maximum likelihood
estimates of g from (4). We make the following regularity assumptions:
a. The vector of attributes z has a distribution p in the
population which has a bounded support.b. The MNL specification (1) with a parameter vector g* is the
true model
.
c. The parameter vector 3* is asymptotically identified, i.e., if
3* 3*, there exists a set Z of z values and an alternative i
such that
2 P(i|z, C, 3*) du (z) * I ^ P(i|z, C, 3) du (z). Under these
assumptions, E S. = P(i|z , C, 3*)) the log likelihood converges uniformly
in 3 to
plim L (3) = / .2_ P(i|z, C, 3*) In P(i|z, C, 3) du (z), (1.5)
and (5) has a unique maximum at 3 = 3*. Then the maximum likelihood
estimator 3_ is consistent, and /n (3-3*) converges in distribution to a
normal random vector with zero mean and covariance matrix
plim (-3^L ( 3*) / 9693'). Discussion and proofs of these properties can be
found in Manski and McFadden (1981); see also McFadden (1973).
Let A = {1,...,m} be a subset of the choice set C. Consider the
conditional log likelihood of the subsample who make choices from A. If the
MNL specification (1) is true, then the IIA property states that the
probability of choosing i from C, given that the choice is contained in A,
equals the probability of choosing i from A. The conditional log likelihood
is then
L,(3) = -^ "L. .E^ S. In (P(i|z", C, 3)/P(A|z", C, 3)) (1.6)A N n=l leA m
= ^ Z, .E^ S. In P(i|z'^, A, 3).N n=l leA in '
Some components of 3*, such as the coefficients of alternative-specific
9
variables for excluded alternatives, are not identified by choice from A.
Let z = (y ,x ) be a partition of the explanatory variables into a vector y
which only varies outside A and a vector x which varies within A, and let
6 = (y>9) be a conmensurate partition of the parameter vector. The
conditional choice probability is then
n„ n„X.9 X. 8
f.
P(i|x^ A, 8) = e ^ / .E . J ^^-'^
jeA
and y. = y. for i, j e A. We add to the regularity assumptions the
asymptotic identification condition
d. If e?' 9*, there exists a set Z of z values and analternative ieA such that
f P(i|x, A, e*) du(z) t (^ P(i|x, A, 6) d M (z).
Then, as in the unconditional case, the conditional log likelihood converges
uniformly in 9 to
plim L^( 9) = /^E^ P(i|z, C, 3*) In P(i|x, A, g) du (z)^^ g^
= / P(A|z, C, e*) .E P(i|x, A, 9*) In P(i|x, A, 9) du (z)
with a unique maximum at 9 = 9*, The maximum likelihood estimator 9. is^ A
consistent and Tn" (9 - 9*) is asymptotically normal with mean zero and
covariance matrix plim (- 3^1 ( 9*) / 8939' )
.
The specification test statistic is based on the parameter difference
6=9-9, where 3 = (y > 9 )• When the regularity assumptions hold and
the MNL model is true, plim 6=0. Conversely, when the MNL specification
10
(1) is false, then the IIA property fails, and (7) becomes
plim L^(9) = / P(aU, C, 3*).E^ [ p(^|^; g; gj ] in P(i|x, A 0) dy (z)
with P(i|z, C, 3*) ^ P(A|z, C, B*) * P(i|x, A, 6*). In general, equation
(1.8) is not maximized at 9 = 6*, implying plim 6^0. Thus, a test of 6 =
is a test of the MNL specification. Rejection of 5 = indicates a failure
of the restrictive structure of the MNL form embodied in the IIA property, or
a misspecification of the explanatory variables z in (1), or both.
Acceptance of 6=0 implies that for the given specification of explanatory
variables and distribution of these variables, the IIA property holds. Thus
the test is consistent against this family of alternatives. However it is
not necessarily consistent against all members of the family of alternatives
defined by a given specification of explanatory variables and any
distribution of these variables.
To derive an asymptotic test statistic for 6=0, note that under the
regularity assumptions, vN (6-3*, 9. - 9*) is asymptotically normal with
mean zero and a covariance matrix V calculated below; the argument is a
standard application of a central limit theorem, as used in Manski and
McFadden, (1981).
We require the gradients and hessians of the likelihood functions and
their moments. First, the moments of S- aremE(S^^) = P(i|z", C, 3*) (1.10)
11
(p(i|z",C, 3*)(1-P(i|z", C, e*)) if i=j
"''^^in'^jn^ "|-P(i|z",C, 3*) P(jlz", C 3*) if i^jj (l.Il)
The gradients are
N
8L^(3)/83 = - I, -^^ S. 31n P(i|z", C, 3)/93
i nil iic(^in-^^^l^"'^'^)^(^i-^C^ (l-^^^
where
and
z"" = .E^ z" PCilz'', C, 3) (1.13)C ^ "
, N3L,(e)/8e = 4 ^, -^A s. 8inp(i|x", A, e)/3eA N n=l leA m
= i I, ^, is. - S, PCilx"", A, e)) (x"- x") f. ...N n=l leA in An i A (1.14)
where
n „ n , . , n „.x, = .E^ X. P(i X , A, 6)
,
A leA 1
S. = .E, S. (1.15)An leA m
The hessian for the unconditional log likelihood is
H E - 32 L (6)/33 33' = - E 32l (3)/3333'
N
=-J:
E, .E^ P(i|z", C, 3)(z" - z")(z" - z")'. (1.16)Nn=l].cC lAiAFor the conditional log likelihood,
1N
-32l fe)/3e39' = i- E, -E^ S, P(i|x",A.e)(x" - x;!)(x" - x")' (1.17)A N n=l leC An i C i C
and hence
12
H, E - E 92 L,(e)/3e3e'^
, N ^
= 4 ^1 -^A P(A|z", C, 3*)P(i|x", A, e)(x'? - x")(x" - x")' (1.18)N n=l leA i A i A
Evaluated at 6 = 6*. H. is thenA
Ni v^ r. /-iH „,\/n n.,n n.,
(1.19)
We now turn to calculation of the asymptotic covariance matrix V.
Taylor's expansions of the gradients imply
H„ >rN(B^-3*)
/N(e^-9*)
y/F 3L^(3*)/3B
/n 3L,(e*)/3BA
+ A (1.20)
where A is a vector satisfying plim A = 0. Using the expectations of the
outer products of the gradients of the two likelihood functions we calculate
iN( 3L ( e*)/33' , 3L (6*)/38') E g' is asymptotically normal with mean zero
and covariance matrix
lim E g g' = lim
H,
0'
-Jc
(1.21)
where H is the H from equation (1.16) evaluated at 3=i
*H. is H. evaluated at 0=6*.A A
and likewise
From (20) , the asymptotic covariance matrix is
- i
V = lim
H,
H,
(limEgg') lim
-1
13
= limH H
-1C "C
Define V. = lira H. ^ andA „ A
Cyy CyGVCOy
VC89
where the partition is commensurate with (Yj9). Then
V,
V =
V VCyy CYe Cy9
C9yV Vcqy cee
C06 C(
V.
The asymptotic covariance matrix of N (
Q = ^ - ^cee'
,) is then
the difference of the asymptotic covariance matrices of 9 and 9 .
test statistic
T = N(9^ - 0^)' Q (9^ - 9^),
(1.22)
(1.23)
(1.24)
(1.25)
Thus the
(1.26)
where Q is a generalized inverse of Q, is asymptotically distributed chi-
square with degrees of freedom equal to the rank of Q. This statistic then
coincides with the general specification test statistic developed by Hausman
(1978) and generalized to use of singular covariance matrices by Hausman-
Taylor (1981) when an efficient estimator is available under the null
hypothesis.
The estimated covariance matrices for the maximum likelihood estimator
3 and 9 satisfyt» A
14
GOV (Q ) =C
/ii „ , s , n n.,n n.2,, .L P(i z , C,3„)(z. - z„)(z. - z)n=lLEC Ci Ci C
-JL
ue
plim N cov O ) = V
cov ( e )
A
Nn n, , n
I. S, .Z, P(ilx", A, e,)(x'.' - x'')(x'.' - x';)'n=l An ieA ' A i A i A
-1
plim N cov ( 9 ) = VA A
(1.27)
(1.28)
(1.29)
(1.30)
Therefore, an asymptotically equivalent computational formula for T is
itT = (6^ - 0^)' [cov (e^) - cov (9^)] (e^ - 6^) (1.31)
which should be quite easy to calculate using existing logit programs.
The appropriate degrees of freedom can be computed using
df = tr [cov (6^) - cov (9^)]*^ [cov (6^) - cov (0^)]. (1.32)
The regularity assumptions do not exclude the possibility that Q is less than
full rank; however, deficiencies will occur only for exceptional
configurations of the x. variables. In particular, a sufficient condition
for Q to be non-singular is that
Hp«o - h" = ^ Z, .E^ P(i|z'',C,3*)(x'? - x")(x'' - x")'C90 A Nn=lieD i Di D
n n,, n n.+ ^ J^ P(A|z",C,B*) P(D|z';c,B-")(x;;-xp(x^-x^)' (1.33)
have a non-singular limit, where D = C|A and
This is generally the case if the x variables either vary within D, or
15
take on values within D different from their average within A.
The estimated matrix [cov (9.) - cov ( 9p) ] may fail to be definite in
finite samples even when Q is non-singular. This does not impede calculation
of the statistic (3A) or carrying out the asymptotic test. However, one can
form an asymptotically equivalent estimate of cov (9.) such that
[cov (6 ) - cov (9 )] is always positive semidefinite by evaluating
P(i|x , A, 9) in (32) at 9 and replacing S by P(A|z ,C,e ). We have\j An L»
occasionally found the test statistic of equation (1.32) to be negative due
to lack of positive seraidef initeness in finite sample applications.
Replacement by the alternative covariance matrix always leads to a small
positive number. However, in no case have we found this alternative
statistic to be so large as to come close to any reasonable critical value
for a X test.
16
II. An Application
As an application of the test, we consider consumer choice of clothes
dryers. The alternatives are an electric dryer, a gas dryer, or no dryer.
We consider a three-alternative MNL model of this choice. A plausible
alternative is that the two types of dryers are viewed as having many common
unobserved characteristics in the decision whether to have a dryer. Hence,
we apply the specification test by excluding the no-dryer alternative. The
sample is 1408 households from the 1975 WCMS survey of residential appliance
holdings and energy use patterns; detailed discussions of the sample can be
found in Newman and Day (1975) and Dubin and McFadden (1980).
The variables used in the model are: (1) clothes dryer operating cost
in 1975, (2) clothes dryer capital cost, (3) homeowner dummy variable,
(4) number of persons in the household, (5) gas availability index. The
operating costs were calculated on the assumption that each person in the
household dries two loads of clothes per week. The capital costs we
calculated as the mean price plus a connect charge. The connect charge was
not included for the gas dryer if other gas appliances were already owned.
Similarly, a connect charge for the electric dryer was not included for the
electric dryer if 220 volts using appliances were already owned. It should
be noted that the choice variable reflects holdings rather than purchases.
Our analysis of holdings as a function of 1975 costs requires the strong
assumption that these reflect expected cost at date of purchase. We assume
in addition that the choice is made by the resident as an independent
17
decision, and not by a landlord or as part of an overall dwelling purchase
decision. Finally, while the GASAV variable and capital costs reflect the
probabilities that all alternatives are in fact available to the decision-
makers, we are unable to identify those individuals in the sample who are
captive to one of the alternatives and unable to exercise choice. These
limitations of the data and the specification of the explanatory variables
make it important to emphasize that our test is a joint test of the IIA
structure of the MNL model and of the variable specifications, and a
rejection may be due to either or both.
In the first column of Table 2.1 we present estimates of the MNL model
for the three choice (unrestricted) situation. We have included choice
specific dummy variales for the electric dryer and no dryer options with a
normalization that the gas dryer dummy variable equals zero. Therefore for
instance, homeowners are less likely to buy electric dryers than gas dryers
and are even less likely still to have no dryer. Also the availability of
gas raises the probability quite strongly that a gas dryer will be owned.
Both operating cost and capital cost coefficients have the expected negative
sign, and they are estimated quite precisely. Their estimated ratio is .81
which leads to a calculated discount rate of about 120% per year assuming a
lifetime for dryers of 8 years. This discount rate seems decidedly too high.
We now re-estimate the model after deletion of the no dryer alternative.
Thus we compare the choice between an electric dryer and gas dryer among the
subset of households that have a dryer. If the IIA property holds true then
the coefficient estimates should remain approximately the same. Under the
alternative hypothesis, we might expect the coefficients to change because
18
gas dryers and electric dryers are closer substitutes than the no dryer
choice. Estimates from the restricted choice set are reported in the second
column of Table 2. Note that while the capital cost coefficient remains the
same that the operating cost coefficient has almost tripled. The estimated
discount rate is now about 40% which is high but in line with other estimates
for appliance purchase decisions, e.g., Dubin and McFadden (1981) and Hausman
(1979). The other coefficients which change markedly are the effects of the
number of persons in the household. To calculate the specification test we
use equation (1.31) where the dimension of 6 is 6 since the 4 no dryer
specific variables are eliminated. We find T = 20.9. Since under the null
hypothesis T is distributed as central x with 6 degrees of freedom, we
reject the MNL specification at beyond the 99 percent critical level.
In the last column of Table 2.1 we present an alternative Hausraan-type
specification test where we have eliminated the electric dryer option to form
the restricted choice set. Since the unrestricted logit estimates are
calculated from the weighted average of the binary choice set estimates we
might expect the estimates from the gas dryer-no dryer choice set to also
provide the basis for a test. Examination of the results in Table 2.1
indicates that the estimates differ greatly from the unrestricted estimates.
In fact, the coefficient of operating cost now has the incorrect sign. We
2find T = 94.1. Again under the null hypothesis T is distributed as x with 6
degrees of freedom. Therefore, we reject the MNL specification at beyond the
99 percent critical level. Note, however, that the two specification tests
of this section are not independent as the combined size of the two
2tests is not l-( .99) . We return to this subject subsequently. Since both
19
specification tests lead to such a decisive rejection of the MNL
specification for the dryer choice problem, we may safely conclude that
misspecification is a serious problem with our model. We now consider an
alternative model specification to the MNL, the nested logit model which
allows for more flexible patterns of similarities. We also develop an
alternative specification test based on classical testing procedures. In
cases where the nested logit model is the correct specification, it might be
expected to provide a more powerful test from the omnibus-type Hausraan test
which in this application is not based on a specific parametric alternative
hypothesis
.
20
TABLE 2.1
Unrestricted and Restricted MNL Choice Models
Variable
1. Operating cost
2. Capital Cost
3. House owner x
Elec. dryer dummy
4. Persons x elec.Dryer dummy
5. Gas availabilityX Elec . dryer dummy
6. Elec. dryer dummy
7. House owner xNo dryer dummy
8. Persons x
No dryer dummy
9. Gas availability x
No dryer dummy
10. No. dryer dummy
Log Likelihood
Number of observations
AlternativeUnrestricted Restricted Restricted
Estimate Estimate Estimate(A.S.E.) (A.S.E.)
-.036
(A.S.E.)
-.013 .195
(.005) (.007) (.038)
-.016 -.016 -.073
(.001) (.001) (.012)
-.631 -.695(.280) (.298) —
.068 .227
(.056) (.076) —
-1.229 -1.696(.476) (.507) —
2.096 2.616(.458) (.494) —
-1.691 -1.88(.263) — (.282)
-.376 .117
(.049 — (.090)
-1.512 -.536(.497) — (.609)
.093 -14.4(.558) (2.97)
-1344.9 -498.8 -407.6
1408 961 809
21
III An Alternative Nested Logit Specification and Classical Tests
The use of the Hausraan-type specification test in Sections I-II requires
no specific alternative model. ^ In this section we consider a specific
alternative model, the nested logit model, to base test procedures on. Given
a parametric alternative hypothesis, we can apply classical test procedures
such as the Wald test, likelihood ratio (LR) test, and Lagrange multiplier
(LM) test. It is well known that for local deviations from the null
hypothesis, these tests have certain optimal large sample power properties,
c.f. Silvey (1970) and Cox and Hinckley (1974). Of course, the optimum
properties of these classical tests depend on three factors which may not be
satisfied in a given application: (1) the alternative specification on which
these tests are based is correct, (2) the sample is large enough so that the
asymptotic theory provides a good approximation, (3) deviations from the null
hypothesis model are of order l/VN. For our dryer choice application of the
previous section, it appears that the nested logit model provides an
attractive model for the alternative specification. In many other
applications the choice of the alternative model may not be nearly so
clearcut. The question of sufficiently large samples and local deviations we
postpone to the next section.
^It is interesting to note that the previous test is not equivalent to anANCOVA-like procedure in which 3 would be allowed to vary across each
alternative or some subset of alternatives after a normalization. The
specification test here would involve a test of the equality of the 3's. But
it is straightforward to check that the IIA property still holds under this
specification so that the most likely failing of the MNL model would not be
tested
.
22
The nested logit model for the three choice case has the simple
hierarchial nature shown in Figure 3.1. Alternatives 1 and 2 are assumed to
have more common characteristics than either alternative has with alternative
3.
Fig. 3.1
The idea behind the choice process is that the individual forms a weighted
average of the attributes of alternatives 1 and 2, sometimes called the
inclusive value which is closely related to his consumer's surplus from these
two choices considered above.
= log (e"-'^'^
+e ^2^/^)(3.1)
where X is a scalar parameter of the model. The choice probabilities of the
model are
P(l|z,C,|3,A) = p =
z,3/X Xye ^ e
I e y (e"3e^ e^y)
p(2|z,c,e,x) = P2 =ZpS/X Xy
e ^ e__
y . Z3B Xy.e (e ^ + e ^)
(3. 2. a)
(3.2.b)
23
P(3lz,C,3,X) = P3 = —-r —
-
(3.3.b)e
For A = 1, the nested logit model reduces to 3 MNL model. For 0<A<1 the
model fails to satisfy the IIA property but it does satisfy the properties
required for a random utility model. This proposition is proven together
with a discussion of other features of the general model specification in
McFadden (1981). One property of the model which deserves mention is that on
any subbranch of the tree, the ua assumption is made. However, in our
application since only a binary comparison involving alternatives 1 and 2 is
required, the IIA property does not lead to a testable restriction.
Given the alternative the classical test procedures may be applied by
basing tests on the likelihood function
N
1n=l Aec
IN
L.,(3,X)=- I Is. log P(i|z",C,3,X) (3.4)N N ^, , in
The assumptions for the MNL logit model following equation (1.4) are
sufficient to prove consistency and asymptotic normality of the estimates
(3,X) = 6 where we replace 3 by 6 in the assumptions. The covariance matrix
of the asymptotic normal distribution equals plim (- B^L ( 6*)/ 3686' )
.
The classical tests proceed so that each test leads to a test of A=l, so that
24
under the null hypothesis the test statistic is a central x^ random variable
with 1 degree of freedom. ^ But a potential problem arises in that the
likelihood function of equation (3.4) is not inexpensive to maximize as
application of both the Wald test and LR test require. A ridge in the
likelihood space makes for slow progress for most algorithms. The likelihood
function is not globally concave as is the MNL likelihood function. An
alternative procedure would be to maximize equation (3.4) conditional on X
and to choose the maximized value of L with respect to both g and X.N
However, each of these conditional maximizations requires an optimization
routine so that the entire process is not inexpensive. Since we want to find
an inexpensive test procedure for the MNL model, we instead based our tests on
one Berndt-Hall-Hall-Hausraan (1974) step beginning from consistent parameter
estimates (3,X). At least since R.A. Fisher's 1925 article, it has been
known that the resulting estimates are asymptotically equivalent to the ML
estimates .^
We derive our intial consistent estimates from the so-called "sequential
estimator", McFadden (1981). The sequential estimator first estimates the
vector g/A and an estimate of the inclusive value y from estimation based on
the restricted choice set, A which contains alternatives 1 and 2. The
estimate of A is then derived by use of a MNL program which considers the
^Here we take the alternative hypothesis to be A*l. One might consider themore restricted alternative that A<1 given the random utility modelrequirements. Since only 1 parameter is under test, the critical values for
the test could be easily adjusted.
^Fisher considered the i.i.d. case while here the presence of the z^'s leadto a i.n.i.d. case. However, the proof of asymptotic equivalence is
straightforward
.
25
choice between alternative 1 or 2 and alternative 3, along with the estimate
B Now in theory the BHHH step leads to asymptotically efficient estimates;
in practice, we found it was a common occurence for the likelihood function
to decrease relative to its value at the consistent, but inefficient,
sequential estimates.
Therefore, we took two or more steps until an increase in the value of
the LF occurred. The asymptotic theory remains the same although the
results are somewhat worrisome in light of our sample size of nearly 1500.
The results of these procedures are given in Table 3.1. In the first
column we chose the 'correct' nested model where the two dryer choices are
combined on one branch of the tree. For our nested logit specification
alternatives 1 and 2 are electric and gas dryers with alternative 3 being the
no dryer option. The estimates of g change significantly from Table 2.1.
The implied discount rate is now about 72%, in between the restricted and
unrestricted logit estimates. Gas availability is now estimated to be less
important, but the no dryer dummy variable has a much larger effect. In
terms of our test procedures we find the following:
(1) The estimated extra parameter is X = .364 which is 5.17 asymptotic
standard deviations from the MNL value of 1.0. The Wald test is therefore
calculated to be 26.7 which gives a very strong rejection of the null
hypothesis. (2) The LR test takes twice the difference of the log likelihood
functions and equals 6.8. The MNL specification is again strongly rejected.
26
TABLE 3.1
Nested Logit Estimates and Classical Test Results
Variable
Gas-ElectricNested Logit
Estimate(A.S.E.)
Gas-No DryerNested Logit
Estimate(A.S.E.)
1. Operating Cost
2. Capital Cost
-.0089(.0045)
-.0065(.0020)
-.134(.004)
-.146(.007)
3. House owner xElec. dryer dummy
4. Persons xElec . dryer duirnny
5. Gas availabilityX Elec . dryer dummy
6. Elec. dryer dummy
-.284(.154)
.047
(.035)
-.556(.292)
.944(.381)
.383(.148)
.387
(.041)
-.730(.271)
-2.35(.253)
7. House owner xNo dryer dummy
8. Persons xNo dryer dummy
9. Gas availabilityX No dryer dummy
10. No dryer dummy
-1.50(.197)
-.327(.037)
-.919(.441)
1.01
(.397)
-.763(.061)
.053(.013)
-.256(.119)
-5.62(.319)
11. X
Log Likelihood
Number of obstructions
.364
(.123)
-1341.5
1408
.405
(.025)
-3943.8
1408
27
(3) The LM test calculates ( 9L/ 95) 'Q"
H
3L/ 96) where Q is a consistent
estimate of the covariance matrix of the asymptotic distribution of N (5-6)
under the null hypothesis. Here it is calculated to be 5.65. Thus, all
three classical tests lead to a rejection of the MNL specification. They
also lead to a rejection at a higher level of significance than the Hausman
tests although those tests did reject at beyond the 1% level.
To further investigate the performance of the classical tests, in column
2 of Table 3.1 we present estimates of an 'incorrect' nested logit
specification. Here we made the gas dryer choice correspond to choice 1 in
Figure 3.1 and the no dryer option correspond to choice 2. On the other
branch, the electric dryer option is choice 3. Note that the coefficient
estimates change markedly, especially for those coefficients which interact
with choice specific dummy variables. The calculated discount rate is now over
100 percent since the estimated coefficient of capital cost exceeds the
estimated coefficient of operating cost. The results of the classical tests
are: (I) The estimated parameter is X = .405 which is 23.8 asymptotic
standard deviations from the MNL value of 1.0. The Wald test is therefore
528.9 which is a stronger rejection of the MNL model than the 'correct'
nested logit model gave. (2) The LR test turns out to be unuseable. A
negative LR statistic is calculated even after three iterations at the
initial "consistent" estimates. (3) The LM test is calculated to be 30.9.
Thus both the Wald test and LM tests based on the 'misspecif ied' nested logit
model lead to stronger rejection of the original MNL specification than do
the corresponding tests based on the "correct" nested logit model. As with
the Hausman specification tests, the two sets of tests are not independent.
We next investigate whether the tests can be distinguished on the basis of
28
size and power characteristics. For instance, if the classical tests are
found to have a large power advantage over the Hausman test, they might be
the tests of choice despite their increased computational requirements.
29
IV. Exact and Approximate Comparisons of theHausman Test and the Classical Tests
The results of this section may prove surprising to devotees of
classical test procedures, especially recent advocates of LM tests.
^
We consider a three choice example when the Hausman test and classical tests
have closed form solutions. Exact comparisons of the size and power of the
tests can then be made. To make the comparisons, we choose a nested logit
model as the correct model under the alternative hypothesis. Therefore, the
Wald, LR, and LM tests are based on the correct alternative specification.
The Hausman specification test is, of course, not based on a specific
alternative model. Just these type of conditions lead to theorems of optimal
asymptotic power among the Wald, LR, and LM - the so-called holy trinity - of
asymptotic tests. For our particular example we choose the sample size equal
to 1000 so that asymptotic theory should provide a reasonable guide. Yet we
find two rather unexpected results: (1) The three tests which comprise the
trinity differ markedly in both of their operating characteristics even when
their nominal size is the same.^
^Two classes of LM tests seem worthwhile to distinguish here. One class is
based on initial consistent estimates so that the LM test is equivalent to
taking one Newton type step which under general regularity conditions willproduce asymptotically efficient estimates. The other class of LM tests is
based on initially inconsistent estimates under the alternative hypothesis,and except in the linear case or for local alternatives their finite sampleproperties have not proven to be generally attractive. It is the secondclass of LM tests that we consider in our application.
^That the actual size differs in a predictable way for a class of linearmultivariate regression models was proven by Berndt-Savin [1978].
30
(2) Uncorrected for size, the Wald test and Hausman test have approximately
equal power. Next comes the LR test with the LM test distinctly in the rear.
Both results may well arise because we are considering a non-local
alternative hypothesis.^ That is, the expansions required for the optimal
power theorems hold for parameter vectors 9. = 6 +6 where 6 must beAcsufficiently small, e.g., 6 = d/jN where d is a constant vector. But,
examples have been given such as Peers (1971) where significant differences
can arise. No generally accepted theory exists for the non-local case. Or,
the samples may not be large enough for the reliable application of
asymptotic theory. We see our results as a particular example and as a
caution against relying too heavily on the local asymptotic theory.
For our example we consider a 3 choice MNL specification of equation
(1.1) with only a single explanatory variable. Furthermore, we assume only a
single data configuration occurs, Z|=l, Z2=0, Z3=0. We assume N=1000
repetitions of the choice and cell counts n ,n ,n . We assume that the true
parameter 3=log 2 generates the observations under the null hypothesis.
For these data the MNL choice probabilities take the form:
p^ = e^/(2 + e^) P2 = P3 = ^/^^ + e^) (4.1)
which for 3=log 2 are p = .5 and p = p = .25. The log likelihood function
for the unrestricted choice set is
L^(6) = n^ log pj^ + n^ log p^ + n^ log p^= n R-N log (2 + e^). (4.2)
Maximization of the likelihood function yields the estimates
A discussion of local and non-local alternatives for an application of theHausman test is given in Hausman [1978].
31
^C" ^^g ^ n2 ^ ng ^
V^ = N/n,(n2 + nj) (4.3)
where V is the large sample estimator of the variance of 3 given in equation
(1.16).
For the restricted choice set we eliminate choice 3. Maximization of
the restricted log likelihood function L ( 3) of equation (1.6) yields
estimates
\ = log (—J V^ = (n, ^ n2)/n,n2^^^^^
where V is calculated from equation (1.17). The Hausman test statisticA
based on the deletion of alternative 3 is
2n2
H. = (B -3,)2/(y V ) = do ( ))2 n2(n2+ n3)/n3. (4.5)3 CA CA n2+n3 ^. ^ o o
By symmetry the Hausman statistic based on the deletion of alternative 2 is
2n3Ht = (log ( ))2 n2(n2 + n3)/n3. : (4.6)^ n2+n3 ^^ -^ ^
Note that the last possible test Hjl is not defined under our data
configuration since B is not identified when alternative 1 is deleted.
As the model specification for the alternative hypothesis we use the
nested logit model. For our uses it can be most convenientally written as
Bz3 Bz3 Bzjl/X BZ2/A^
Pi = II|P P2 = n2P P3 = 1-p = e /(e . (e e ) ) (4.7)
Bz ./A Bzi/A BZ2/Xwhere 11. = e /(e + e ) for i = 1,2. The log likelihood has the
form
L(a,X) = n^a - (n|+n2)(l-X) log (e^+l) - N log (^(e^+l)^ (4.8)
32
where we use the parameterization a = 3/ A. The maximum likelihood estimates
"1 "3 n2
a = log X = log (-—-_)/log (-—:—
)
n2 n i+n2 n i+n(4.9)
\.^^2
Denote the partitioned large sample estimate of the information matrix as
-lim E— — —
N-»-oo L \x- A
aa \x_\x hx_ Aa hx_ (4.10)
Then the Wald statistic for the hypothesis H : A=l is
W = ( X - 1) 2 V"
^
n 2n 3
where V~ ^ = A,,-A^jA = —:— t?( t ?'^t|n3/N)/( t ?n2/(n ,+n2) + t§n3/N)AAaAaa N ii^o i^ i^ <io
for ti = log (n2/(n|+n2)) and t2 = log (n 3/(n i+n 2) ) .^
(4.11)
Next the LR statistic is calculated from the unrestricted log
likelihood function of equation (4.2) for the MNL model which sets A=l and
the nested log likelihood of equation (4.8):
2n. 2n3LR = 2(L(a,A)-L (3)) = 2n2 log ( ) + 2n3 log ( ) (4.12)
c "^ n 2+n 3 -^ n 2+n 3
Finally we derive the LM statistic. Under the null hypothesis with A=l, we
^Details of these derivations will be provided by the authors uponrequest
.
33
use the MNL estimate 3 = log (2n|/(n2+n3)) so that in equation (4.7)
II2 = (n2+n3)/(N+ni ) and p = (N+nx)/2N. We then evaluate the gradient of the
likelihood function (4.8) at this point to find
n 3~n2
L. = ( —7 )log (n2+n3)/(N+njt) (4.13)A Z
together with its large sample variance
V"-^ = - (EL,,- (EL, )2/EL ) (/ M\\\ Aa aa (4.14;
3N-ni (n2+n3)(N+nj,)
InTF: ^N(log((n2+n3)/(N+n,)))2.
Therefore we calculate the LM statistic as
LM = l2/V~^ = N(n3-n2)2(3N+nx)/(3N-ni)(n2+n3)(N+ni) (4.15)A
Asjmiptot ically, each of the test statistics H3, H2, W, LR, LM is under
the null hypothesis distributed x?- Note that (nx,n2,n3) has the trinomial
distribution,
N! n| n2 n3
Pr (njL,n2,n3) = —-j—^—y- pj^ p2 P3 ,(4.16)
where Px,P2>P3 ^re given by equation (4.1) with A=l under the null
hypothesis, and for 0<A<1 under alternatives. For the example, we calculate
numerically the exact distribution of the statistics for X=l and for
alternative values A = ( .95 , . 9, .85 , .8 , . 75 , . 7 , .6) . These calculations permit
determination of the exact sizes of each test for various nominal sizes, and
of the power functions, either size-corrected or uncorrected. The means and
34
variances of the parameter estimates and test statistics, conditioned on
positive cell counts, are also computed.
In general when closed form test statistics cannot be found, numerical
calculation of exact power requires expensive Monte Carlo simulation. A
"rough-snd-ready" approximation to power can however be calculated as
follows: Let S(ni,n2,n3) be the statistic in question, and let
U = S(Npi,Np2,Np3), (4.17)
where (Pi»P2>P3) ^^^ the choice probabilities under an alternative, be a
measure of the "location" of the statistic. If S is asymptotically x^ under
the null hypothesis, then under alternative S is roughly a non-central chi-square with noi
centrality parameter ij , and power can be calculated from the distribution
kT + m
P (x'2 =< X) = e-'^' y^-^^"(^-/^^
y(-^/2)-
m=0 m!(4+ra) n=o"-^"'"'- r(-^+n). (4.18)
Table 4.1 gives the exact distribution of the five alternative test
statistics under the null hypothesis: HAUS 3 is the Hausman statistic based
on deletion of alternative 3, HAUS 2 the corresponding statistic for deletion
of alternative 2, and WAl.D, LM, and LR are the Wald, Lagrange multiplier, and
Likelihood ratio statistics, respectively. Table 4.2 gives the means and
variances of these statistics, exact sizes of various tests, and critical
levels.
Table '+.1 shov,s that the exact distributions of the test statistics are
relatively close to their asymptotic limit. All the statistics have
thicker lower tails than the chi square. The Wald test also has a thicker
35
upper tail, and the LM test a thinner upper tail, than the chi square where
we use the upper tail to define the critical point for the tests. The
conditional means and variances given in Table 4.2 also indicate that the
Hausraan statistics and likelihood ratio statistic have distributions close to
the asymptotic limit, while the Wald statistic is distributed with a higher
mean and variance and the LM statistic with a lower mean and variance. The
exact sizes of the LR, HAUS 3, and HAUS 2 tests are close to their nominal
sizes when the asymptotic critical level is used. However, the exact size
36
TABLE 4.1
Exact Cumulat ive Distribuition Function of the Test Statistics
iARGUMENTCHISQUARE HAUS 3 ILAUS 2 WALD LM LR
.0001562 .0100000 .0178485 .0178485 .0178485 .0178485 .0178485
.0009766 .0250000 .0178485 .0178485 .0178485 .0178485 .0178485
.0039119 .0500000 .0535097 .0535097 .0535097 .0535097 .0535097
.0157204 .1000000 .0890643 .0890643 .0890642 .0898139 .0890643
.0356288 .1500000 .1595404 .1595404 .1551770 .1595720 .1595490
.0639752 .2000000 .1943906 .1943906 .194 3880 .1986097 .1943891
.1012512 .2500000 .2571302 .2571302 .2333807 .2625366 .2574044
.1481302 .3000000 .2964287 .2964287 .2956866 .3088109 .2963552
.2055111 .3500000 .3533769 .3533769 .3334747 .3615748 .3547693
.2745780 .4000000 .3964012 .3964012 .3910138 .4110970 .3948666
.3568926 .4500000 .4514412 .4514412 .4352908 .4667838 .4527005
.4545310 .5000000 .5000545 .5000545 .4849164 .5174321 .5005644
.5702946 .5500000 .5501913 .5501913 .5342529 .5663164 .5490352
.7080494 .6000000 .6011664 .6011664 .5828786 .6169712 .5987006
.8732966 .6500000 .6506719 .6506719 .6327198 .6672093 .64904551 .0741902 .7000000 .6995021 .6995021 .6832754 .7179841 .70083061 .3235021 .7500000 .7501350 .7501350 .7341925 .7682552 .75001171 .6428286 .8000000 .8004872 .8004782 .7847090 .8171284 .80030992 .0730254 .8500000 .8504647 .8504645 .8360386 .8652888 .8501056.7067207 .9000000 ,9001828 .9001823 .8881028 .9120306 .9004682
3 .8431482 .9500000 .9499366 .9499344 .9419375 .9570080 .95032315 .0258684 .9750000 .9748535 .9748486 .9696143 .9782958 .97541366 .6369923 .9900000 .9898452 .9898323 .9870385 .9907046 .9904561
10 .8291046 .9990000 .9989388 .9989182 .9984343 .9985983 .999210315 .1371548 .9999000 .9998936 .9998833 .9998088 .9997743 .999960019 .5106637 .9999900 .9999924 .9999901 .9999792 .9999684 .999999923 .9262119 .9999990 1 .0000000 .9999998 .9999989 .9999975 1 .000000028 .3710278 .9999999 1 .0000000 1.0000000 1 .0000000 ]L. 0000000 1 .0000000
37
TABLE 4.2
Characteristics of the Test Statistics
CHISQUARE HAUS 3 HAUS 2 WALD LM LR
Mean^ 1.0 0.99990 1.00009 1.07102 0.92832 0.99554
Variance^ 2.0 2.01212 2.01608 2.30900 1.69288 1.95064
Exact sizeof a nominal10% test^
0.10 0.09982 0.09982 0.11190 0.08797 0.09953
Exact sizeof a nominal
5% test
0.05 0.05006 0.05007 0.05806 0.04299 0.04968
Exact size
of a nominal1% test^
0.01 0.01015 0.01017 0.01296 0.00929 0.00954
Critical level 3.84315 3.84528 3.84535 4.11628 3.57253 3.8231J
for an exact
5% test.
^ Mean and variance are conditioned on the event that all cell counts
exceed 50, which occurs with probability 1-10"^^.
2 Critical level 2.70672.
3 Critical level 5.02587.
38
of the wald test exceeds its nominal size substantially, while the exact size
of the LM test is substantially lower .^ For instance, for a nominal 5% test
size, the Hausman tests are within 1% of the nominal size and the LR test is
within .6%. But the Wald test size is off by 15.8% and the LM test by 15.1%
in the other direction. For a nominal 1% test the discrepancy of the Wald
test is 26%. Table 4.2 gives the adjustments in critical levels required to
achieve an exact 5 percent test with each statistic.
Table 4.3 gives the power of each of the tests against the nested logit
model with values of X less than one. Power curves are given for both the
asymptotic tests and the size corrected 5 percent tests. Also given are the
rough estimations to power based on equations (4.17) and (4.18).
The exact powers of the asymptotic tests, uncorrected for size, are
ranked
WALD > HAUS 2 > LR > LM > HAUS 3
over most values of X. The performance of the WALD and LM tests is explained
in part by the deviations between the exact and nominal sizes of these tests.
The differences in power are of sufficient size and uniformly to suggest that
the WALD and HAUS 2 tests are clearly preferable to the remainder. The LR
and LM tests do not do nearly as well by comparison.
Table 4.3, III gives the powers of the size corrected tests. Here the
rankings are
^These sizes results for the WALD, LR, and LM tests are in accord with theBerndt-Savin [1978] results for multivariate linear models.
39
TABLE 4.3
Exact Power of the Tests
I. Asymptotic test, nominal size 0.10
Alternative X
1.0 0.95 0.90 0.85 0.80 0.75 0.70
HAUS 3
exact .0998 .1447 .3437 .6078 .8434 .9604 .9950
approx. .0999 .1554 .3534 .5993 .8190 .9402 .9871
HAUS 2
exact .0998 .1685 .3893 .6544 .8714 .9700 .9966
approx. .0999 .1583 .3817 .6659 .8937 .9824 .9989
WALDexact .1119 .1829 .4091 .6728 .8817 .9733 .9971
approx. .0999 .1624 .3992 .6911 .9103 .9870 .9994
LMexact .0879 .1416 .3445 .6094 .8445 .9608 .9950approx. .0999 .1531 .3507 .6056 .8349 .9541 .9929
LRexact .0995 .1564 .3669 .6320 .8583 .9656 .9959approx. .0999 .1568 .3670 .6316 .8579 .9648 .9955
40
Table 4.3 (continued)
II. Asjrmptotic test, nominal size = 0.05
Alternative X
1.0 .95 ,90 .85 .75
HAUS 3
HAUS 2
WALD
LM
LR
0501 .0778 .2278 .4730 .7475 .9210 .9869
0500 .0885 .2438 .4745 .7245 .8928 .9722
0501 .1015 .2826 .5419 .7998 .9438 .99190500 .0906 .2682 .5451 .8242 .9633 .9971
0581 .1130 .3029 .5650 .8158 .9501 .9931
0500 .0936 .2835 .5728 .8479 .9719 .9981
0429 .0779 .2332 .4807 .7540 .9240 .9876
0500 .0869 .2415 .4810 .7449 .9148 .9837
0497 .0894 .2557 .5090 .7757 .9336 .9897
0500 .0895 .2555 .5082 .7744 .9324 .9891
41
Table 4.3 (continued)
HAUS 3
HAUS 2
WALD
LM
LR
irrected Test, exact size = 0.05
^
Alternative X
1.0 .95 .90 .85 .80 .75 .70
0500 .0775 .2272 .4722 .7471 .9208 .9868
0499 .0884 .2436 .4723 .7243 .8927 .9722
0500 .1015 .2826 .5418 .7997 .9437 .9918
0499 .0905 .2697 .5449 .8241 .9632 .9971
0500 .1014 .2823 .5413 .7992 .9435 .99180425 .0821 .2607 .5458 .8312 .9672 .9977
0498 .0896 .2562 .5098 .7764 .9340 .9898
0588 .0992 .2641 .5091 .7669 .9252 .9864
0502 .0901 .2570 .5107 .7770 .9342 .9899
0506 .0904 .2571 .5103 .7760 .9931 .9892
42
Table 4.3 (continued)
IV. Asymptotic test, nominal size 0.01
Alternative X
1.0 ,95 .90 .85 .80 .75 .70
HAUS 3
HAUS 2
WALD
LM
LR
.0102 .0166 .0755 .2248 .4911 .7646 .9366
.0134 .0208 .0967 .2475 .4924 .7343 .9028
.0102 .0321 .1283 .3251 .6111 .8479 .9669
.0134 .0215 .1104 .3064 .6243 .8797 .9838
.0130 .0381 .1450 .3529 .6396 .8648 .9720
.0134 .0225 .1193 .3316 .6600 .9022 .9888
.0072 .0188 .0856 .2458 .5187 .7855 .9450
.0134 .0203 .0954 .2526 .5174 .7747 .9359
.0095 .0235 .1007 .2753 .5548 .8112 .9545
.0134 .0212 .1032 ..2 748 .5552 .8100 .9533
43
HAUS 2 > WALD > LR > LM > HAUS 3.
In these rankings, HAUS 2 and WALD have comparable power, and are
substantially better than the LR and LM tests, which have comparable power.
The high exact power of the HAUS 2 test is surprising in view of the
fact that the Wald , LM and LR tests are specific against the alternative
model generating the observations. This outcome may be special to this
example where the Hausraan tests have the same degrees of freedom as the
trinity. However, the result also indicates that for this problem the sample
size of 1000 is not sufficient for the asymptotic equivalency of the trinity
and second order efficiency of the LR test to take effect in the sense that
the largest deviation A. = (1.7)/v^= .009 is not sufficiently local for the
asymptotic power rankings to apply.
The rough approximation to power is within .003 of the exact power of
all these tests, and the ranking of the tests by approxirate power usually
coincides with the exact power ranking. We conclude that the rough power
calculation is potentially a very useful tool for quick comparisons of tests
or judgments on small sample power, without the great expense of Monte Carlo
calculations
.
For comparison purposes, we then recalculated the exact classical tests
based on a misspecified nested logit model. That is, in eqution (4.7) we
interchanged p and p so that choices 1 and 3 now lie on the same branch.
However, we continue to conduct the test procedures based on the original
model specification where choices 1 and 2 were on the same branch of Figure
3.1. In Table 4.4 we present both approximate and exact size corrected test
43a
TABLE 4.4
Approximate and Exact Test Results at Test Size .05
Wrong Model Correct Model
X = .95 X = .9 X = .85 X = .95 X = .90 X =
Nominal size 0,05
HAUS 3 .102 .283 .542 .078 .228 .473
HAUS 2 .078 .228 .473 .102 .283 .542
WALD .089 .250 .501 .113 .303 .565
LM .078 .233 .481 .078 .233 .481
LR .089 .256 .509 .089 .256 .509
Exact size .05
HAUS 3 .101 .283 .542 .077 .227 .472
HAUS 2 .077 .227 .472 .101 .283 .542
WALD .078 .228 .473 .101 .282 .541
LM .090 .256 .510 .090 .256 .510
LR .090 .257 .511 .090 .257 .511
MLE estimate 1.050 1.112 1,171of X
44
results for the Hausraan specification test and classical test procedures.
The two Hausman tests interchange operating characteristics since our example
is symmetric with respect to them. Therefore, HAUS 3 is now the more
powerful test. The Wald test which had excellent operating characteristics
when based on the correct model now loses substantial power. Note that the
Wald test is now based on inconsistent estimates. The power characteristics
of the LR and LM tests are symmetric with respect to the models. Thus, the
HAUS 2 is now significantly more powerful than the classical tests. The
ranking based on exact sizes is
HAUS 3 > LR > LM > WALD > HAUS 2
The HAUS 3 has approximately 10% more power than the LR and LM tests and 15%
more power from the Wald test. Thus, it seems useful to investigate
different tree structures when using the classical tests because of their
sensitivity to model specifications. In the last line of Table 4.4 we also
present the maximum likelihood estimates of X which all exceed the theoretical
maximum of 1.0 which again indicates misspecification , Likewise, it seems
useful to base the Hausman specification test on different restricted choice
sets since its power also depends on the model specification.
45
V. Conclusions
The finding that the Hausraan test has better power than the LR or LM
test for our example may make further evidence on the test statistics of
interest. We also engaged in very limited Monte Carlo examples on the
original data for the application of Sections 2 and 3. We took the
unrestricted MNL estimates of Table 2.1 to be the true values for the g's and
then generated choice outcomes using the nested logit model of equation (3.2)
to be the true model. For a range of X's we simulated choice behavior via
the generalized extreme value distribution, McFadden (1981),
F(.e^,e^,e^] = exp [-exp ([(-e^ + (-e^) ] - e^)] (5.1)
and then assigned the choice for each individual to the alternative with
the highest utility calculated by the random utility model, u.=z.6 for j=l,3. We
then compared the Hausman test of Section 2 with 6 degrees of freedom to the
trinity of tests from Section 3 which all have 1 degree of freedom. The
values of X that we chose were X = ( .25 , .5 , .75 , .9 , .95 , 1 .0) . At the 5% level
all tests rejected for X = ( .25 , . 5 , . 75 ) . For X = (.9, .95,1.0) none of the
tests rejected. As in our example of the previous section the Wald test
always achieved the highest value (after we normalized the Hausman test) when
it was based on the correctly specified model. But we cannot correct for
size here since we do not have exact values. Additionally, we generated
samples using the multinomial probit model of Hausman-Wise (1978).
46
A somewhat disturbing result here is that while the Hausraan test does not
reject the independent probit specification, which is quite similar to MNL
with IIA except for differences in the extreme tails of the underlying
distributions, the trinity of test rejects resoundingly with the statistics
above 5.0. We do not yet have a satisfactory explanation for this outcome.
In terms of applying the Hausraan test or the trinity of tests, the non-
uniqueness of application can arise. For instance in the 3 choice case any of
the 3 alternatives can be dropped or 3 different tree structures can be
defined. As the number of choices grows, the different possible combinations
grow factorially. If more than one test is performed, the problem of
controlling for the size arises because the tests will not be independent.
In the case of the Hausman test the different estimates could be combined to
form a test after their joint asymptotic covariance matrix is calculated
which would not be difficult^. Alternatively, using the IIA property of the
null hypothesis, the restricted choice set A could be successively decreased
when more than 3 choices exist and the size controlled for as we do when
independent F tests in linear models are used. However, we have no
prescription about optimal procedures in any given application.
We conclude that the Hausman test and Wald test are the best choices to
test the IIA assumption in MNL models. The Wald test does require maximum
likelihood, or at least asymptotically efficient estimates, of the nested
logit model and correct specification of the tree structure in the nested
^We attempted such a combination for our example in Section 4. But for that
example we found the tests to have perfect negative correlation. Therefore,it appears that it is possible to increase the power indefinitely for anynominal size test by choosing the appropriate weighted average of the test
statistics. We leave to further investigation the problem of optimalcombination of the tests.
47
logit model. The Wald test has higher power in all our examples than the LR
or LM test for the correct specification. However, for an incorrect
specification the LR and LM tests are superior to the Wald test. But the
Wald test has significantly greater computational requirements than does the
Hausman test. An alternative test based on a consistent estimate of X from
the sequential logit estimator is possible by the use of Neyman's [1959] C
procedure. However, our experience with the sequential logit estimator is
that it gave quite unreliable estimates of X method for testing.^ Since the
Hausman test gave results in general close to that of the Wald test across
the range of our examples, we recommend it as a general purpose specification
test for the MNL model. The Wald test should also be considered when the
analyst feels that the nested logit model provides the correct specification
for the choice problem under consideration. The Wald test requires more
sophisticated computer software; furthermore, once it is size corrected it
no longer is superior in our example to the Hausman test which is very
accurate with respect to its size. Furthermore, for the case of a
misspecified nested logit model, the alternative Hausman test performed much
better than the Wald test. But certainly some test of the IIA property
should be made when the MNL model is used. Our experience is that the
Hausman test has rejected the MNL specification in a number of applications.
^An alternative indication of the unreliability is that in the majority ofthe cases, one BHHH step beginning at the sequential logit estimates led to a
decrease in the value of the likelihood function. Since the C test is baseda
on the one step methodology, we decided against its use. Furthermore, in the
misspecified model case, we were often unable to find an increase in thelikelihood function even when up to three BHHH steps were made.
48
REFERENCES
Berndt, E., B. Hall, R. Hall, and J. Hausman (1974), "Estimation andInference in Nonlinear Structural Models", Annals of Economic and SocialMeasurement, 4.
Berndt, E. and G. Savin (1977), "Conflict Among Criteria for TestingHypotheses in the Multivariate Linear Regression Model", Econometrica,45.
Cox, D.R. and D. Hinckley (1974), Theoretical Statistics , London: Allen andUnwin.
Debreu, G. (1960), "Review of R. Luce, Individual Choice Behavior", AmericanEconomic Review, 50.
Domencich, T. and D. McFadden (1975), Urban Travel Demand, Amsterdam: NorthHolland.
Dubin, J. and D. McFadden (1980), "An Econometric Analysis of ResidentialElectrical Appliance Holdings and Usage", working paper, Department of
Economics, M.I.T.
Hausman, J. (1978), "Specification Tests in Econometrics", Econometrica 46 .
Hausman, J. (1979), "Individual Discount Rates and the Purchase and
Utilization of Energy Using Durables", Bell Journal of Economics , 10.
Hausman, J. and W. Taylor (1981), "The Relationship between Classical Testsand Specification Tests", MIT mimeo
.
Hausman, J. and D. Wise (1978), "A Conditional Probit Model for QualitativeChoice", Econometrica , 46.
Manski, C. and D. McFadden (1981), "Alternative Estimators and Sample Designsfor Discrete Choice Analysis" in C. Manski and D. McFadden (eds)
Structural Analysis of Discrete Data , Cambridge: MIT Press.
McFadden, D. (1973), "Conditional Logit Analysis of Qualitative ChoiceBehavior" in P. Zarembka (ed.) Frontiers in Econometrics, New York:
Academic Press.
McFadden, D., W. Tye, and K. Train (1976), "An Application of DiagnosticTests for the Independence from Irrelevant Alternatives Property of the
Multinomial Logit Model", Transportation Research Record 637.
49
McFadden, D. (1981), "Econometric Models of Probabilistic Choice", in C.
Manski and D. McFadden (eds), Structural Analysis of Discrete Data,
Cambridge, MIT Press.
Newman, D. , and D. Day (1975), The American Energy Consumer , Cambridge:Ballinger
.
Neyman, J. (1959), "Optimal Asymptotic Tests of Composite StatisticalHypotheses", in U. Grenander, ed
., Probability and Statistics, Stockholm :
Alraquist and Wiksell.
Peers, H. (1971), "Likelihood Rates and Associated Test Criteria",Biometrika, 58 .
Silvey, S. (1970), Statistical Inference, London: Penguin.
Date Due
JAN 18 19B9
MAR 3 V ]Q(.
JULoSlfeo
ACMEBOOKBINDING CO., INC.
SEP 1 5 i933
100 CAMBRIDGE STREET
CHARLESTOWN, MASS.
ARIES DUPL
[j'^iDflD DDM ms Tbl
2^(
3 TDflD DDM M77 3fl3
•
MIT LIBRARIES _DUPL __' -^Q
J<^3
[_3 TDfiD DD4^ MIS 17^
3 TDflo DD^ ms Ta7MIT LIBRARIES
3 TDflD DDM 415 TTS
MIT LIBRARIES .^93
3 =^060 DD4 41b DPIMIT LIBRARIE5__ ,.,,, 1111 nil II =i'^V
3 TDfiD DDM M Mb "133
I
^-f
3 TDflD DDM M7fl DSDmiT LIBRARIES '^f'^
3 TDSD DDM MMb "IMl